In order to provide for fully automated analysis of BGAs, software is being developed, that identifies, analyses and classifies the balls of BGAs.
Since algorithms for ball counting and void detection have already been implemented and are already detailed in the public domain emphasis of this research has been on the analysis of the wetting of the balls.
Pre-processing
Due to short exposure times, the X-ray images are noisy. In a first step this noise is reduced by a rank filter using a moving 5x5 square window. In the next step the objects are segmented from the background. The goal is a binary visualisation, where black pixels denote objects like BGA-balls and white denotes background.
However, simple segmentation using a global threshold for binary visualisation fails due to variations in the background as can be seen in Figure 2. Here, a global threshold was set to separate the balls on the right side. The balls on the left side are still mixed with the background. If a global threshold for the balls on the left side is chosen, the balls on the right side would get smaller and change shape.
In order to facilitate accurate segmentation, thresholds are computed locally. In the newly developed algorithm a mask is moved across the image in steps. At each step, the histogram of the grey values inside the mask is analysed and a threshold is determined that applies for the whole mask. According to this threshold the image is binary visualised.
In the histogram analysis the threshold is determined by finding a value that on the one hand separates the dark and light values evenly and on the other hand, defines a local minimum (see figure 3).
Fig.1 X-ray image of corner of BGA
Fig.2 Corresponding segmented image using global threshold
Fig.3 Segmented image using variable threshold
Ball Separation
Depending on the viewing angle, adjacent balls may not appear separated sufficiently well in the image or even overlap. Up to a certain degree, it is possible to separate those balls by image processing. First, small holes and gaps are filled by applying a closing operation. Then the balls are separated by applying a distance transformation followed by a watershed transformation. Figure 4 shows a magnified part of an X-ray image before and after the separation step.
(a)
(b)
Classification
For each ball, a number of features are extracted. To name a few, the area, angle, height and width are measured. Also the curvature and roughness of the contour of each ball is analysed. The features of well wetted balls are used to create a model of "good" balls. Then the individual objects are classified according to their features (Green:- pass, Red:- fail, Yellow:- borderline, blue:- cannot process reliably).
Fig. 5(a) X-ray acquired image of
BGA balls
Fig. 5(b) Classification of
the wetting of BGA balls for image shown in fig. 5(a)
The developed software allows accurate segmentation of the objects from the background. This is a prerequisite for many defect detection algorithms like ball counting or void detection. An algorithm was developed that analyses the wetting of the balls by classifying the balls according to several features like area, angle and contour-based features. The software was tested with X-Ray images of BGAs with artificially introduced poor wetting. The balls with poor wetting were reliably detected. Currently the software is being tested with many real-world images of BGAs with poor wetted balls to verify if slight adjustments of the classification and possibly of the extracted features are necessary.
